RerankRunPipeline

Overview

The RerankRunPipeline class is made for using reranker easily. To use this class, you can use Reranker easily without extra work. For further information about reranker, please check out Reranker documentation.

In this pipeline, first retrieval retrieves passages. Then reranker reranks retrieved passages. Finally, put retrieved passages to LLM model.

Usage

Initialize

To create an instance of RerankRunPipeline, you need to provide an instance of a Retrieval class and Reranker class. Optionally, you can specify which LLM model you want to use.

from RAGchain.pipeline import RerankRunPipeline
from RAGchain.retrieval import BM25Retrieval
from RAGchain.reranker import MonoT5Reranker

retrieval = BM25Retrieval(save_path="path/to/your/bm25/save_path") 
reranker = MonoT5Reranker()
pipeline = RerankRunPipeline(retrieval=retrieval, reranker=reranker, llm=OpenAI()) 

Run Pipeline

The run variable is runnable of Langchain LCEL. So you can use all method for running pipeline, like invoke, stream, batch, and so on.

answer = pipeline.run.invoke({"question": "Where is the capital of Korea?"})

If you want to get used passages or relevance scores of retrieved passages, you can use get_passages_and_run method.

answers, passages, scores = pipeline.get_passages_and_run(["Where is the capital of Korea?"])